Google Gemini 4 Argon Is Here. We Tested It (+11 AI Updates)
Publicado no IA Dots ·
Agentes de IAAI SignalAssistente pessoaldot-anthropicdot-artificial_intelligencedot-claudedot-geminidot-googledot-metadot-microsoftdot-musedot-openaiidioma-enInglêsMuseyt-2XpIMa1BSPc
Vídeo em Inglês, publicado pelo canal Vaibhav Sisinty.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
This video covers multiple major artificial intelligence updates, including the release of Google's Gemini 4 Argon, Meta's Muse Wonder, OpenAI's GPT-6 with intelligent UI, Anthropic's Claude Haiku 5.5 and anti-distillation safeguards, Microsoft's Copilot PC integration and Maya 300 chip, Sarvam AI's Bulbul V4 Flash, Alibaba's QuenJoy, and various tests comparing Gemini 4 Argon and Claude.
Key points
- Google launched Gemini 4 Argon with a 1 million token context window, advanced reasoning, and availability across multiple Google platforms.
- Meta introduced Muse Wonder, a personal AI model featuring active memory and active focus to help users achieve long-term goals.
- OpenAI released GPT-6 with intelligent UI, allowing the model to generate interactive visuals and tools directly in conversations.
- Anthropic implemented safeguards against distillation attacks after catching AI companies using Claude to improve their own models, alongside launching Claude Haiku 5.5 and Claude dashboards/motion.
- Microsoft updated Copilot with hybrid intelligence to manage PC files and tasks, and unveiled the Maya 300 AI chip.
- Practical tests compared Gemini 4 Argon and Claude on tasks such as building interactive Rubik's cubes, retro games, video editing projects, and 3D concept car websites.
Topics
Gemini 4 ArgonMuse WonderGPT-6Claude Haiku 5.5Copilot PC IntegrationMaya 300 ChipModel BenchmarksAI Agents
Video transcript
Transcript from available captions; it may contain errors.
Google is back [music] and its new Gemini 4 argon could give Claude a serious run for its money. But does it actually live up to the hype? We tested both on realworld tasks. At the end, you'll see which perform better and how to try Gemini yourself. Meanwhile, Mark Zuckerberg just took a major step towards personal super intelligence. Our vision to build [music] personal super intelligence for everyone. Meta's new Muse Wonder can remember what matters to you [music] and is designed to work towards your goals for months, even when you're not using it. And OpenAI just changed how chat GPT works. [music] Instead of giving you long written answers, GPT6 can now create interactive visuals you can actually use. But while companies are racing to build smarter AI, Anthropic has caught other AI labs using Claude to improve their own models, and it's fighting back. [music] Those are just four of this week's 12 biggest AI updates. And at the end, we'll put Google's biggest claims to the test and show you exactly how to use Gemini 4 Argon yourself, step by step. As always, if any section gets too technical, you can jump straight to the next [music] update you care about. Everything is timestamped in the description. I've also put all the prompts, tools, and useful resources inside our [music] free WhatsApp community. The link is in the description. Now, let's get into the video. Google just made its most advanced AI model for complex reasoning available worldwide. It's called Gemini 4 Argon, and it comes with a massive 1 million token context window. You can use it to analyze complex documents, improve your code, create 3D visualizations, and even turn your ideas into interactive experiences. And the best part, Gemini 4 Argon is now available through the Gemini app, Google search, Google Anti-gravity, [music] and the Gemini API depending on your plan. Now, we were so impressed by what this model can do that we decided to test it ourselves. And later in this video, we'll show you exactly how to use it with a complete step-by-step tutorial. But before we get to that, let's go through the other biggest AI updates of the week. Meta just launched Muse Wonder, its most powerful AI model yet. And the goal is to build an AI assistant that understands your life, remembers what matters to you, and helps you get things done. One of its biggest features is active memory, which means you don't have to explain everything from scratch every time. For example, if your washing machine breaks down, you can ask Muse Wonder to find your warranty and help you file a claim. or if you're helping your father recover after knee surgery, it can help you keep track of his recovery schedule. And then there's active focus. When you give Muse Wonder a difficult task, it's designed to keep looking for solutions even when it runs into problems. Meta claims it can even continue working towards your goals in the background for days, weeks, or months. And that's Meta's bigger vision, personal super intelligence for everyone. an AI that remembers what matters to you and helps you achieve your goals. OpenAI just changed how Chat GPT answers your questions. It launched GPT6 with a new feature called intelligent UI. And now ChatGpt can show you interactive visuals instead of just giving you long paragraphs of text. For example, imagine you're trying to assemble a bicycle. Instead of going through a complicated instruction manual, Chad GPT can show you an interactive 3D model so you can see exactly how the different parts fit together. Or let's say you're exploring San Francisco. Chat GPT can show you an interactive map with places you can visit. You can even preview different paint colors on your walls before actually painting your room. And look at this. Chat GPT can also guide you through making an origami fox with step-by-step animations instead of just written instructions. And it goes beyond visuals. Chat GPT can now create interactive tools like calculators, charts, and buttons directly inside your conversation. There's another useful upgrade. GPT6 can now start showing you an answer while it's still working on your question. So you don't have to wait for the entire response. And this is rolling out to everyone, including free users. So instead of just reading chat GPT's answers, you can now interact with them and actually see how things work. Anthropic just caught AI companies using Claude to make their own model smarter. It identified activity involving Chinese AI companies Z.A.I. and Moonshot which were using Claude's capabilities to improve their own AI models. This technique is known as a distillation attack. Now, Anthropic has introduced new safeguards to prevent this. And here's what's interesting. According to Anthropic, these defenses have reduced the amount of information attackers can extract by 90 to 99%. Anthropic has also banned tens of thousands of accounts linked to these attacks. So, the AI race isn't just about building smarter models anymore. It's also about protecting them from being copied. Microsoft just gave Copilot a major upgrade that lets it work directly on your Windows computer. With hybrid intelligence, Copilot can now access your files, organize documents, and complete tasks on your PC with your permission. For example, it can search through your files, rename them, and move them into the right folders automatically. And it can handle more than just organizing files. You can simply ask Copilot, I just got an email from my accountant, can you help me with my taxes and get her what she needs? C-pilot searches your computer, finds the required documents, organizes everything into one folder, and drafts an email to your accountant. And before sending it, C-pilot asks for your approval. Also, Copilot can now run AI models directly on your computer instead of sending every task to the cloud. This can make certain tasks faster, reduce costs, and help keep sensitive data on your device. And when it needs more power, it can use online AI models, too. So instead of just answering questions, Copilot can now work with your files and complete tasks directly on your PC. Microsoft just unveiled a powerful new AI chip to reduce its dependence on NVIDIA. It's called Maya 300 and it's designed to power AI tools like Copilot and Microsoft Teams. According to Microsoft, it can produce 40% more AI output with the same amount of electricity as leading GPUbased systems. And with 96 chips working together, it can deliver up to seven times the output of the previous generation. Microsoft plans to start deploying Maya 300 in 2027. So Microsoft is now building its own chips to make AI faster, cheaper, and more efficient at a massive scale. India is taking on the US in AI voice technology. SAVAM AI just launched Bul V4 Flash, a new AI voice model that can generate surprisingly natural and expressive speech in multiple Indian languages. It can deliver cricket commentary in English, health conversations in Canada, storytelling in Marathi, payment reminders in Odia and ads in Hindi. It can change its tone and emotion to match the situation and even mix Indian languages with English. Naturally, Bulbull V4 flash is now available through Saram's API with a limited time launch discount. Anthropic just launched its cheapest and fastest clawed model yet, its clawed haiku 5.5. And the biggest surprise is how much cheaper it is. According to Anthropic, Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average while being much more capable. And the improvements go beyond just speed and cost. Haiku 5.5 is now much better at coding, using computers, and handling tasks that involve multiple steps. It's also designed to handle repetitive work like sorting large amounts of information and creating summaries. And here's what's interesting. You can now control how much effort cla puts into each task. So for simpler jobs, you can keep it fast and cheap while harder tasks can get more attention. Anthropic has also cut the cost of reusing previously processed information with claude sonet 5.5 by 50%. Haiku 5.5 is already available to developers through Anthropic, Amazon, Google and Microsoft. So, Anthropic is making AI not just more capable, but much cheaper for companies that need to run thousands of tasks every day. Claude can now turn your company data into interactive dashboards and even animated videos. Anthropic just launched Claude dashboards and Claude Motion. For example, say you run a bike rental company and every morning certain stations keep running out of bikes. To figure out why, you connect your company's data to Claude. Claude goes through an entire year of ride data and creates a dashboard showing when people rent bikes, which stations are busiest, and how many bikes are left. Now, if you want to explain what's happening at your next team meeting, you can simply select a chart and ask Claude to turn it into an animation. Claude then creates a video showing how bikes move across the city throughout the morning. You can even leave comments directly on the video, like, "Show the empty bike station a little longer." And Claude makes the change. Your teammates can also leave comments and you can export the finished animation into your design tools. So instead of walking your team through complicated charts, you can now show them exactly what's happening through a simple animated video. Google just launched a new AI agent that can work like an entire team inside your company. It's called the one agent and unlike a regular chatbot, it can understand your work, connect to your apps, and keep working 24/7 even when you're not there. You can upload a business proposal request and simply ask Gemini to complete it. Gemini read the document, figures out what needs to be done, creates a plan, and asks for approval before starting. Once approved, Gemini creates multiple AI agents to handle different parts of the job. One checks customer records, another searches the web, and another reviews company requirements. All these agents work together while Gemini coordinates everything, and the work continues in the background. The next morning, the completed draft is ready. You can use it across tools like Gmail, Google Drive, Slack, and other business apps. Google also lets you choose between different AI models, including Gemini and Claude depending on the task. And companies get security controls to decide what the AI can access and which actions need human approval. So, instead of giving AI instructions for every small task, you can now assign Gemini an entire job and let it coordinate the work for you. Anthropic just launched a new initiative to protect some of the world's most important systems from cyber attacks. It's called the Anthropic Cyber Mission and it focuses on two major areas. First, Anthropic is working with companies like Crowdstrike, Deote, and Palo Alto networks to help secure power grids, water systems, factories, and transportation networks. Claude will help security teams find and fix weaknesses before attackers can exploit them. Second, Anthropic launched OSS Scanner, a free tool that regularly checks open-source software for security problems. It can identify potential weaknesses, explain how they could be exploited, and suggest fixes. The reports are AI generated, so they still need verification. So, Anthropic is now using Claude to help protect the infrastructure and software that people depend on every day. Alibaba has found a way to train AI agents without wasting as much computing power. It's called Quenia. Think of it as a smarter way to manage the computers that teach AI agents how to complete complex tasks. When an AI agent is training, it may need different amounts of computing power at different stages. Quenia can shift those resources around while training is still happening instead of letting them go to waste. It also helps reduce repeated work when an agent explores different ways to solve a problem. In Alibaba's tests, this approach cut training time by around 44% and improve performance on a coding benchmark. In simple terms, Quenjaya helps AI agents learn more efficiently using less time and computing power. Google is back with Gemini 4 Argon, but can it actually beat Claude? Google claims this is one of its most powerful AI models yet with major improvements in coding, reasoning, and handling complicated tasks. And if you look at Google's demonstrations, the results are pretty impressive. But we've seen this before. Every new AI model launches with amazing demos and benchmark scores. And almost every company claims its model is the best. So what happens when you give these models the same real world tasks? That's exactly what we're going to find out. We tested Gemini 4 Argon on several projects and gave Claude the same instructions to see how their results compare. And we're not just checking whether they can complete a task. We're looking at the design functionality, small details, and whether you'd actually want to use what they create because a model can look incredible on a benchmark and still struggle when you give it real work. By the end, you'll know where Gemini stands against Claude and which model makes more sense for your work. I'll also share all the prompts inside our free staying ahead WhatsApp community. The link is in the description. But before we get into our test, Google has shown something with Gemini 4 Argon that's worth taking a closer look at. Google shows a humanoid robot learning to play cornhole, a game where you throw bean bags into a hole in a board. Researchers first showed the robot a video of a person making the throw. Gemini studies the movement, then uses the robot's cameras to work out exactly where the board is. Next, it practices inside Mujoko, a simulator that lets it test different movements before trying them in the real world. After every throw, it checks where the bag landed, looks at what went wrong, and adjust the next attempt. In Google's demonstration, the robot eventually lands seven shots in a row. The interesting part is how Gemini combines what it sees in a video with practice and feedback. And Google has shown the same model creating some very different visual projects. On Google's website, you can see codegenerated 3D pagodas, a racing track, a chessboard, a room, and detailed Gothic architecture. Its model card also covers engineering, finance, legal work, and cyber security. And it supports up to 1 million tokens across input and output, giving it room to work with a huge amount of information. The benchmark results are strong. Although Gemini doesn't lead every test, including Frontier SWE and Terminal Bench in Google's comparison, but instead of spending all our time on charts, let's open the tool and test the actual results. For these projects, we're using Google's anti-gravity workspace. Open the model menu, select Gemini 4 Argon, and you'll see low, medium, and high effort settings. We used high for the Gemini tests, so let's begin with a project that needs both good design and working controls. First, we asked Gemini to build an interactive Rubik's cube. Let's open the finished project called Cube Studio Pro. You can switch between 2x two, 3x3, 4x4, 5x5, and mirror cube puzzles. There's also a finish menu with options like brushed gold, carbon fiber, and cyber rows. Now, let's click scramble, then try autosolve. You can open a flat 2D view, check the move history, and choose preset patterns. It even has an exploded view that separates the pieces plus controls for zoom and animation speed. So you have several different ways to play with the cube. Now let's open Claude's version of the same task. You get different cube sizes. And if you look at the textures and listen to the sound effects, the cube feels much more realistic. You can also scramble and solve it. Track your moves and time and switch between silver and gold finishes for the mirror cube. So, Gemini gives you more customization options, while Claude's version looks and feels more polished, but can they build something that plays without you controlling it? Next, we asked Gemini to create a Chrome style dinosaur game that plays itself with a retro Windows theme. Let's run the existing project locally. It's called Cyber Rex Pixels, and you can see the dinosaur moving through a colorful pixel art environment. On the right, there's a task manager style panel with a neural network graphic, tuning controls, and logs showing the game's decisions. You can increase the speed up to 16x, switch between autopilot, and manual mode, and adjust the music, sound effects, and colors. Now, let's open Claude's version, Nebula Runner. It has a darker space theme, automatic play, a score counter, and speed settings. And watch what happens when you close the window. You get standby, turn off, and restart, and selecting restart plays a retro boot animation. Both versions are playable, but Gemini has more game controls, while Claude adds more personality to the interface. Now, let's try something that you might actually use when creating content. We gave Gemini a reference edit and footage of an AI generated presenter talking about Neuralink. The goal was to create a vertical video in a similar editing style. Let's open the finished file. You'll see split screens, animated captions, pop-up text, and motion graphics appearing alongside the narration. Now, let's play Claude's version using the same material. It adds supporting footage, animated text, and transitions that follow what the presenter is saying. Both give you useful first edits. You'd still want to check the timing and details before publishing, but it's interesting to see how much of the editing each model handled. And if you think design matters in a short video, it matters even more on our next project. Let's search for Jaguar Type O and open the official website. This futuristic concept car is our design reference. We asked Gemini to create a scrolling website where a 3D version of the car moves as you scroll. Now, let's open what it built. You get animated sections, a rotating car, an exploded view, and materials information. The scrolling is smooth, but the 3D car doesn't look quite as convincing as the original. Now, let's open Claude's version, also created at high effort. The Blue Jaguar looks more detailed, and as you scroll, you see the wheels, color options, butterfly doors, and exploded view animation. The lighting, typography, and overall design feel more consistent. So, Claude produces the stronger looking website in this test. But before picking a winner for everything, it's worth looking at the cost of getting those results. On the artificial analysis chart shown here, Gemini scores around 53 on the intelligence index, compared with around 58 for the clawed opus model shown. There's also a speed comparison, although the chart doesn't give us a clear comparison between all the models we're discussing. But when you look at cost per task, Gemini is cheaper than several leading models. So, the best choice depends on the quality of the result, how quickly you need it, and what it costs to produce. There's one more test, and this time we're moving out of the browser. For the final project, we connected Gemini to Blender through MCP, which allows the AI to work directly with the software. We use this reference photo of a man with his mouth open and asked Gemini to turn it into a 3D model. Now, let's open Blender and compare the result with the original photo. What's useful is that this is an actual 3D model. You can keep editing, not just a generated picture. You could prepare it for 3D printing, although the recording doesn't show a physical print. So, after all these projects, which model should you actually use after these tests, you can see why it's worth judging models by what they create rather than just what their benchmark pages promise. Gemini handles several kinds of projects while Claude delivers the better looking Jaguar website in this comparison. So instead of choosing one model for everything, compare the results and pick the one that works best for your task. You can find all the prompts and resources in our free staying ahead WhatsApp community. The link is in the description. And if you want to explore more of what the latest AI models can do, watch this video next. Finally, around 70% of people watching these videos still aren't subscribed. So if you want to stay ahead with the latest AI updates and tutorials, hit that subscribe button. I'll see you in the next video.
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, Assistente pessoal, Muse. Publicado no YouTube em 2026-10-15.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.